This is a first step towards #5952. For each third-party distribution, a separate mypy call is started. This has a few advantages, but also a few drawbacks:

• Missing dependencies in METADATA.toml are now spotted, since only the dependencies listed there are being added to the list of files to test.
• It's possible to add a custom configuration for a distribution to METADATA.toml. Previously, all configurations were merged before mypy was run, therefore all configurations applied to all files. Now, only the configurations of a distribution and its dependencies are merged when testing a distribution.
• Adding a non-stub dependency will crash the tests. (But this wouldn't work anyway, at the moment. This is actually a first step towards allowing this.)
• Stubs are now tested multiple times if they are part of the dependency chain of another distribution.

In the future, each distribution should get its own venv, so that the dependencies can be installed from pypi. See #5952 for the rationale.

The code in mypy_test.py is quite janky and hard to read at this point. I wanted to keep refactorings out of this PR, but might send a separate cleanup PR later.